LatentChem: From Textual CoT to Latent Thinking in Chemical Reasoning

Xinwu Ye; Yicheng Mao; Yuxuan Liao; Jia Zhang; Yimeng Liu; Li Hao; Fang Wu; Zhiwei Li; Zehong Wang; Zhiyuan Liu; Zhenfei Yin; Li Yuan; Philip Torr; Huan Sun; xiangxiang Zeng; Mengdi Wang; Le Cong; Shenghua Gao; Xiangru Tang

Source-linked AI summary

LatentChem: From Textual CoT to Latent Thinking in Chemical Reasoning

Xinwu Ye, Yicheng Mao, Yuxuan Liao, Jia Zhang, Yimeng Liu, Li Hao, Fang Wu, Zhiwei Li, Zehong Wang, Zhiyuan Liu, Zhenfei Yin, Li Yuan, Philip Torr, Huan Sun, xiangxiang Zeng, Mengdi Wang, Le Cong, Shenghua Gao, Xiangru Tang

arXiv:2602.07075v6Feb 6, 2026physics.chem-phcs.AIcs.CLcs.LG

TL;DR

Chemical LLMs face a potential modality mismatch when continuous chemical logic is forced into discrete natural-language CoT. LatentChem decouples reasoning from generation with continuous thought vectors and dynamic perception; it reports stronger benchmark performance and lower reasoning overhead, while sacrificing interpretability.

Problem
Chemical reasoning is commonly expressed through explicit CoT, but it remains unclear whether discrete linguistic reasoning is optimal for complex chemical tasks.
Method
LatentChem decouples chemical reasoning from language generation using continuous thought vectors, dynamic molecular refinement, and outcome-driven policy optimization.
Results
59.88% non-tie win rate on ChemCoTBench and 10.84× average reasoning step overhead reduction are reported against explicit CoT across evaluated benchmarks.
Takeaways & Limitations
The study reports spontaneous internalization and presents continuous latent dynamics as a more effective computational substrate for chemical logic than discretized linguistic trajectories.
Takeaways & Limitations
Latent reasoning sacrifices explicit intermediate-step transparency, complicating scientific auditability and trust when reasoning must be inspected.

Abstract

from arXiv · show

Current chemical large language models (LLMs) predominantly rely on explicit Chain-of-Thought (CoT) to solve complex reasoning problems. However, forcing nonverbal tacit chemical logic into discrete natural language imposes a fundamental ``modality mismatch,'' creating an artificial bottleneck for reasoning. We introduce LatentChem, a reasoning interface that decouples chemical logic from linguistic generation, enabling the model to process information via continuous thought vectors and dynamic perception. Our investigation reveals a pivotal emergent behavior: spontaneous internalization, defined here as self-selected under outcome-only optimization. When optimized for task success, the model abandons verbose textual derivations in favor of implicit latent computation, suggesting that it identifies the continuous manifold as a more native substrate for chemical logic. This paradigm shift also proves to be a superior computational strategy: LatentChem achieves a 59.88\% non-tie win rate against the strong CoT baseline on the rigorous ChemCoTBench, while delivering a broad 10.84$\times$ average reduction in reasoning step overhead (5.96$\times$ wall-clock speedup) across all evaluated benchmarks. Our results provide empirical evidence that chemical reasoning is more naturally and effectively realized as continuous latent dynamics rather than discretized linguistic trajectories.

1. Introduction

LatentChem challenges explicit linguistic Chain-of-Thought as the computational substrate for chemical reasoning and introduces a continuous latent interface with dynamic molecular perception. Under outcome-only optimization, it spontaneously internalizes reasoning and achieves both efficiency and performance gains over explicit CoT.

Chemical LLMs typically linearize continuous physicochemical intuitions into discrete natural-language reasoning steps.
LatentChem addresses this modality mismatch by decoupling chemical reasoning from language generation through continuous thought vectors and a dynamic perception loop.
Spontaneous internalization occurs when outcome-only optimization leads the model to discard verbose textual reasoning and collapse computation into latent space.
10.84× average reasoning step overhead reduction and 5.96× wall-clock speedup are achieved across all evaluated benchmarks.
59.88% non-tie win rate against the strong explicit CoT baseline is achieved on ChemCoTBench alongside improved token efficiency.

2. Related Work

Related work spans domain-specific chemical LLMs, reinforcement learning for validity and consistency, and emerging latent-thinking approaches that move reasoning beyond explicit token generation.

Domain-specific chemical LLMs integrate chemical geometries, task decomposition, and reinforcement learning to address specialized scientific reasoning.
Existing chemical systems remain prone to hallucination and often lack robust planning capabilities for complex chemical reasoning.
Latent-thinking research shifts reasoning from explicit token generation to implicit processing in high-dimensional latent spaces.

3. LatentChem: A Latent Reasoning Interface

LatentChem combines molecular alignment, continuous latent thinking, dynamic perceptual refinement, and progressive training to decouple chemical reasoning from linguistic generation.

LatentChem uses a Chemical Adapter, ChemUpdater, and latent projector to connect molecular perception, dynamic refinement, and continuous reasoning.
The Chemical Adapter compresses variable-length molecular features into fixed ChemTokens that condition the language model.
At each latent step, a hidden state becomes a thought vector that drives both recurrent reasoning and active molecular representation updates.
The latent projector feeds a continuous vector back into the backbone until termination or the latent-step budget limit, after which the refined context yields the explicit response.
A four-stage training protocol progresses from molecular-linguistic alignment and explicit CoT supervision to latent activation and GRPO policy refinement.
The outcome-only reward combines format adherence, answer validity, and answer correctness without rewarding brevity or CoT omission.

4. Emergent Properties of Latent Chemical Reasoning

LatentChem develops emergent latent reasoning behaviors during optimization, replacing verbose textual chains with computation in continuous latent states. These states are causally necessary, and the model flexibly shifts between latent and explicit reasoning as its latent budget changes.

Spontaneous CoT Internalization: After reinforcement learning, LatentChem internalized explicit textual reasoning despite prior supervision with explicit CoT data.The transition occurred during Stage 4 optimization.
Spontaneous CoT Internalization: The model replaced intermediate textual steps with latent thinking, a trivial transition artifact, and immediate generation of target XML tags.This pattern was observed under rewards constrained to format adherence, validity, and correctness.
Causal Necessity: Replacing initial latent steps with Gaussian noise caused overall performance degradation, indicating that early latent states encode functionally essential structural precursors.The causal analysis covered molecule optimization and understanding tasks.
Optimization Over Imitation: The internalization emerged from outcome-oriented reward optimization rather than an explicit penalty on output length.The authors interpret this as evidence favoring continuous latent representations over discrete natural-language reasoning for these chemical tasks.
Emergent Compensatory Dynamics: When the latent budget was sufficient at T ≥ 6, CoT length stayed near zero; below T < 6, the model reactivated explicit CoT without external instruction.The budget sweep jointly tracked explicit CoT length and task performance.
Emergent Compensatory Dynamics: The budget response indicates that LatentChem arbitrates between implicit and explicit reasoning pathways according to available computational capacity.The observed pathways behave as communicating vessels rather than a rigid reasoning format.

5. Benchmarking Latent vs. Explicit Paradigms

LatentChem is evaluated across diverse chemical tasks, where it generally outperforms explicit-CoT baselines while substantially reducing reasoning overhead. Gains are strongest for open-ended generation, remain competitive on deterministic mappings, and depend on latent thinking with dynamic perception.

Generative capabilities: LatentChem surpasses all baselines on 14 of 15 metrics for open-ended generative tasks.The evaluation includes molecule optimization and description metrics, with results reported in Table 2.
Generative capabilities: 82% success rate on GSK3-β, versus 67% for explicit CoT, demonstrates a large molecule-optimization advantage.
Robustness on deterministic mappings: LatentChem maintains a leading position on deterministic mappings, but its margin narrows because creative-exploration benefits are more limited.The method preserves competitive parity with explicit baselines, although absolute reaction-prediction accuracy remains modest.
Overall superiority: 59.88% non-tie win rate on ChemCoTBench and 85.26% on ChEBI-20 show robust superiority over the strong explicit baseline.LatentChem also reaches 55.58% on ChemLLMBench and 49.88% on Mol-Instructions, where it maintains competitive parity.
Inference efficiency: 10.84× average reasoning-step reduction and 5.96× wall-clock speedup quantify the efficiency advantage over explicit CoT.Across tasks, reported reasoning-step reductions range from 5.4× to 29.9×; Figure 8 reports the average and wall-clock results.
Ablation study: Removing ChemUpdater causes a 12.0% decline in optimization success, while removing latent thinking causes a precipitous performance drop.The ablation study also identifies the latent projector as indispensable for mapping thought vectors back to the input space.

6. Discussion and Conclusion

The study concludes that decoupling chemical reasoning from linguistic generation provides an efficient substrate for chemical logic, accompanied by spontaneous internalization and strong benchmark results. However, latent reasoning sacrifices interpretability and remains subject to task, data, reward-design, and training-stability limitations.

Discussion and Conclusion: LatentChem decouples reasoning from linguistic generation and uses compact latent computation correlated with physical chemical topology.
Discussion and Conclusion: 59.88% non-tie win rate on ChemCoTBench accompanies a 10.84× reasoning-step reduction and 5.96× wall-clock speedup.
Limitations and future directions: Latent internalization makes intermediate reasoning opaque, complicating scientific auditability and trust when verification requires inspecting the reasoning process.
Limitations and future directions: The study reports task-dependent gains and regressions, shared public-data overlap, no separate RL prompt data, reward-design sensitivity, and GRPO stability as limitations.

Impact Statement

The paper advances chemical reasoning and molecular-design methods with potential scientific applications, but its efficiency analysis is interpretive and depends on simplifying assumptions. Proposed applications therefore require expert review and safety checks because the capabilities may be dual-use.

Impact Statement: LatentChem targets chemical reasoning and molecular design, with potential support for property optimization, drug discovery, and materials science.
Impact Statement: Generated molecules should be treated as research hypotheses requiring domain-expert review and appropriate safety checks before real-world use.
Scope boundary: The reasoning-overhead analysis is an interpretive appendix under simplifying assumptions, not a formal proof or quantitative predictive theory of runtime efficiency.
Scope boundary: The theoretical setup assumes an idealized shared optimal path and non-degenerate representation speed for discrete-approximation analysis.

A.2. The Curvature–Resolution Theorem

The theorem links discrete approximation cost to representation geometry: higher curvature requires finer advances, while rectified latent trajectories permit coarser discretization. The surrounding evaluation framework distinguishes creative generative tasks from deterministic closed-ended tasks.

A.2. The Curvature–Resolution Theorem: High representation-space curvature forces smaller manifold advances to keep chord deviation within tolerance, increasing the required reasoning steps.A rectified representation reduces curvature and permits coarser discretization.
A.2. The Curvature–Resolution Theorem: A near-zero latent curvature removes the curvature term, whereas persistent text-space curvature can produce larger step counts on sufficiently complex tasks.The stated advantage grows with task complexity when latent representation learning reduces curvature.
A.2. The Curvature–Resolution Theorem: ChemCoTBench separates open-ended generative tasks from precise closed-ended tasks requiring unique deterministic answers.This dichotomy is used to examine latent reasoning in creative exploration versus deterministic execution.
A.2. The Curvature–Resolution Theorem: ChemCoTBench covers molecule understanding, editing, optimization, and reaction prediction across 22 subtasks and 1,495 reviewed examples.The examples draw on public molecular databases and patent corpora and were reviewed by LLMs and 13 chemists.
A.2. The Curvature–Resolution Theorem: LatentChem’s training uses counterfactual alignment and strict freezing to ground chemical reasoning in molecular features and activate latent modules.The counterfactual strategy perturbs ChemTokens, while strict freezing updates only thinker and updater parameters during latent activation.

E. Metrics

The evaluation suite combines error, accuracy, similarity, validity, text-overlap, and optimization metrics to measure distinct chemical capabilities. Metrics are tailored to task type, including structural preservation and exact reaction prediction.

E. Metrics: Mean absolute error measures typical prediction deviation and is used for functional-group and ring-counting tasks.Lower MAE indicates smaller average absolute deviations.
E. Metrics: Scaffold Similarity measures Tanimoto similarity between generated and reference molecular scaffolds, with scores ranging from 0 to 1.Higher scores indicate stronger preservation of the core framework, particularly for Murcko scaffolds.
E. Metrics: Accuracy counts correct outcomes, with task-specific criteria for valid molecular edits, ring-system classification, and Top-1 reaction prediction.For reactions, the highest-ranked output must exactly match the ground-truth molecule or molecules.
E. Metrics: Optimization metrics report absolute target-property improvement and the success rate of molecules exceeding a predefined threshold.Mean improvement indicates typical uplift across samples.
E. Metrics: Fingerprint-based Similarities average Morgan, MACCS, and RDKit fingerprint similarities to assess molecular correctness and structural similarity in reaction prediction.The metric is applied to forward prediction and retrosynthesis outputs.
E. Metrics: Text similarity metrics including BLEU, ROUGE, and METEOR evaluate generated molecular descriptions against reference answers.These metrics are used for molecular description generation tasks.

F. Results for Closed-ended Tasks

Closed-ended results are reported for ChemCoTBench, Mol-Instructions, and ChemLLMBench, using task-specific correctness, error, scaffold, and fingerprint metrics. ChemCoTBench results cover molecule understanding, molecule editing, and reaction-related tasks.

F. Results for Closed-ended Tasks: Closed-ended evaluations span ChemCoTBench, Mol-Instructions, and ChemLLMBench, with results organized across Tables 6 and 7.ChemCoTBench includes molecule understanding, molecule editing, and reaction-related tasks; the other benchmarks contain broad reaction subtasks.
F. Results for Closed-ended Tasks: Table 6 reports MAE for functional groups, Scaffold Similarity for Murcko scaffolds, CR% for ring counts, and Top-1 accuracy and FTS for reactions.These metrics cover distinct closed-ended molecule-understanding and reaction-prediction outcomes.
F. Results for Closed-ended Tasks: Table 7 reports Top-1 accuracy and FTS for reaction prediction and correct rate for reagent selection across Mol-Instructions and ChemLLMBench.The table focuses on broad reaction-related subtasks in those benchmarks.

G. Additional Results for Causal Study

Additional analyses examine generated-CoT length under latent-token corruption, output changes under different latent budgets, and per-task overhead reductions. Supplementary tables also restrict benchmark comparisons to selected evaluated subtasks.

G. Additional Results for Causal Study: Figure 11 supplements the causal analysis by measuring generated explicit-CoT length as the first k latent tokens are replaced with Gaussian noise.The appendix provides additional measurements omitted from the main text for space.
G. Additional Results for Causal Study: Figure 13 breaks down reasoning-step overhead reduction by benchmark and task, with task categories sorted by decreasing reduction.The appendix reports reasonably consistent measured reductions across evaluated benchmarks, supporting averaged main-paper results.
G. Additional Results for Causal Study: Supplementary benchmark tables restrict ChemCoTBench comparisons by omitting specified subtasks and report the selected foundational, optimization, and reaction tasks.The omitted tasks include equivalence, retrosynthesis, RCR, NEPP, and MechSel.

K. Inference Latency and Budget Analysis

The evaluation measures model-side latency under controlled inference settings and compares latent-budget, format-adherence, and training-variant outcomes against explicit CoT.

Latency measurement: Latency is measured at batch size 1 with synchronized CUDA events after discarding three warm-up runs per model load.Evaluation uses bfloat16 inference, PyTorch 2.9.0+cu128, CUDA 12.8, and eight NVIDIA H100 GPUs with one model replica per GPU.
Latent-budget analysis: At zero latent budget (T = 0), optimization performance is largely comparable to explicit CoT, while some topology-sensitive tasks benefit from latent budget.The T = 0 condition suggests fallback to textual reasoning on optimization tasks.
Evaluation diagnostics: Format adherence remains above 97% on most benchmarks except for the untuned text-only baseline.The reported diagnostics distinguish reward-hit rates from format-adherence rates across evaluation benchmarks.
Training variants: The frozen-latent LatentChem design outperforms the jointly optimized latent-module Stage-4 variant in overall performance and average reasoning-step overhead reduction.The authors interpret the joint-training weakness primarily as an optimization issue.

N. ChemUpdater Ablation on Topology-Sensitive Tasks

Ablation results link ChemUpdater to topology-sensitive scaffold performance, while separate analyses assess LCF, evaluation overlap, and stability across random seeds.

ChemUpdater ablation: Removing ChemUpdater reduces Murcko scaffold similarity from 0.81 to 0.67 and ring-system accuracy from 97 to 80.The ablation directly tests whether ChemUpdater contributes structure-aware refinement on topology-sensitive scaffold tasks.
LCF ablation: LCF is not essential for the main gains and is treated as an optional grounding regularizer.The Stage-1 counterfactual-alignment ablation reports the full Open/Closed/All breakdown without LCF.
Evaluation scope: Only 8 ChemCoTBench records, or 0.71%, overlap with ChemCoTDataset, while no overlaps are detected on the other benchmarks.Because GRPO uses the ChemCoTDataset prompt pool as SFT, the reported generalization evidence concerns transfer to disjoint evaluation benchmarks rather than unseen RL prompts.
Robustness analysis: Repeated evaluations across seeds 2026–2075 produce stable mean±std non-tie win rates consistent with Table 3.Additional tables provide win/lose/tie breakdowns for ChemCoTBench, Mol-Instructions, ChemLLMBench, and ChEBI-20.